9 May 2011—Last Wednesday, Intel announced a big change to the electronic switches at the heart of its CPUs. Going forward, the firm will be using three-dimensional transistors to take the place of long-used planar devices.
The new transistors—dubbed "tri-gates"—are a variation on the FinFET, a transistor design that substitutes the flat channel through which electrons flow with a 3-D ridge, or fin. Popping the channel out of plane and draping the gate—which switches the transistor on and off—over it will allow Intel to shrink the smallest features in its transistors from 32 nanometers to 22 nm while cutting power consumption in half. This feat would be impossible to do with the transistor design the company had been using.
How did this 3-D design win its way into production? We asked the coinventor of the FinFET, IEEE Fellow Chenming Hu, a professor emeritus at the University of California, Berkeley, how the new transistors got their start, why we need them now, and where they will go from here.
IEEE Spectrum: We’ve been shrinking two-dimensional, or planar, transistors just fine for 50 years. Why are we seeing a switch to three-dimensional FinFETs?
Chenming Hu: I’ll distill the problem with planar transistors to a single point. It all stems from the fact that it is very difficult to turn off a transistor when it’s very small. In other words, you can’t stop the current flowing through the transistor when you don’t want the current to flow.
I’ll use an analogy to explain this. There is a garden hose lying on a soggy lawn, and you want to stop the water from flowing into this lawn. If there’s a long hose, you can call your friends to come in and put 10 pairs of hands down, and you can stop the water. Now imagine you shorten the hose so you cannot even put one palm on it to stop it. Now you shorten it even more, so you can only put one finger on it. It’s impossible to stop.
In the past 10 years, people have dealt with this garden hose problem in various ways, and one way has been to sacrifice power. For 250-nm transistors, the power-supply voltage was 2.5 volts; for 180 nm, it was 1.8 V; for 130 nm, it was 1.3 V. The pattern was very regular until 90 nm, but it reached a limit. Instead of 0.9 V, you know what the industry used? 1.2 V. Even at 45 nm, the industry still used 0.9 V instead of 0.45 V.
IEEE Spectrum: So current is leaking even when the transistors are off. To get around that problem, you have to use a higher voltage to make the difference between on and off more obvious?
Chenming Hu: Exactly. What’s the consequence of that? Power is proportional to the square of the voltage. So if you use twice as high a voltage as the historical trend, your cellphone will consume four times the power. The pain is just too big to keep going that way. We thought planar technology would run out of steam sometime after 25 nm, and it did.
IEEE Spectrum: How do FinFETs help fix the leaky garden hose problem?
Chenming Hu: Remember, the hose is on a soggy, soft lawn. So what if instead of pressing your finger on this garden hose, you pinch your fingers on the two sides of the garden hose? That’s the analogy. The weak point, the soggy lawn, is the silicon substrate. So you really have to do something on both sides so you’re pinching against something firm, and that’s what the FinFET is doing. We should pinch the channel [where electrons flow] on two sides and on top. The more pinching sides, the better.
Pinching the hose will allow us to use a much, much shorter hose. That’s extremely important. Making things small is really the key of making the electronics cheaper, faster, and lower power.
IEEE Spectrum: The idea for FinFETs has been around for a while. How did it all get started?
Chenming Hu: DARPA [the Defense Advanced Research Projects Agency] sent out a request for proposals in 1996 for ideas to develop electronic switches beyond 25 nm. At the time, the industry was using 250-nm transistors, and the general view was that transistors could not be scaled below 100 nm. But my students and I had already been thinking about how to get transistors to scale to 25 nm and beyond.
There was a quick meeting probably lasting only five minutes between myself and two colleagues—Professor Tsu-Jae King Liu and Professor Jeff Bokor. The meeting was short because we already knew what to do.
I was on a flight to a conference in Japan, and I had about 10 hours, so I just wrote down the technical proposal in longhand. I proposed two structures that we’d been thinking about for a while. One was FinFETs, and the other is what we call an ultrathin-body silicon-on-insulator (UTB SOI).
We got the contract in 1997, and that gave us the resources to demonstrate FinFETs experimentally. A young graduate student named Xuejue "Cathy" Huang made the working device, and the team of three professors and 11 students and visiting researchers published it in 1999.
IEEE Spectrum: How did the industry react to the FinFET paper?
Chenming Hu: It was an instant hit. I remember Cathy and I were invited to Intel Santa Clara just a couple of months after the publication, and in that same year, 2000, I was invited to Intel Oregon twice. At the time, people were asking me how long it would take for the idea to get into production. I said about 10 years, so I guess I was off by one.
IEEE Spectrum: Was the attention that you got unusual for a new transistor design?
Chenming Hu: Extremely unusual. We contributed two things: We figured out a way to make the transistor manufacturable, and we showed how this thing could bring us to 25 nm and to 10 nm. We even figured out how to use the FinFET to solve the two top problems plaguing MOSFETs today—random variations of impurity atoms and variations in gate length (roughly the distance from the source to the drain). So we anticipated a lot of emerging problems and showed that FinFETs can solve them. That was really the first time that the industry believed there was life [after 25 nm].
IEEE Spectrum: What have been the challenges in getting FinFETs to market?
Chenming Hu: In production, there are two areas: One is getting the manufacturing variation controlled well. When you add a fin, then you have to make sure the fin’s height and width are uniform. Probably a bigger limitation is the interaction between the transistor and the circuit design. Intel has the benefit of being a vertically integrated company where the designers and technologists work under the same roof.
IEEE Spectrum: What about the other design you proposed to DARPA, the UTB SOI? How does that design work, and how has it been progressing?
Chenming Hu: SOI has a layer of insulator with a thin layer of silicon on top. Instead of a thick regular round hose, UTB is like a thin, flat hose—one of those compact garden hoses that’s easy to roll because it’s almost like tape. You can pinch with one finger from the top, because the hose is very thin, so you don’t have to push down much.
IBM has been making SOI processors for years, and even today it uses about 40-nm-thick silicon. To get to 20-nm transistor size, you have to use very thin, say, 5-nm silicon. When we proposed this, nobody in the world could provide SOI substrates with such thin and uniform silicon film. Five nanometers is very thin; it’s only about 15 silicon atoms or so in thickness. But the amazing thing is, two years ago, an SOI substrate company called Soitec announced that they can make such a thin film, and they have given samples to many companies, including IBM.
IEEE Spectrum: FinFET transistors are the first to market, but could UTB SOI transistors ultimately win out?
Chenming Hu: Starting in 2002, both FinFET and UTB SOI were listed in the International Technology Roadmap for Semiconductors (ITRS). That’s an industry consensus of what’s needed to keep the industry going. They are the only two that are listed as the likely successors of the planar MOSFET.
I think there’s room for both, at least in the short term. UTBs need less manufacturing and design development work than FinFETs, because UTBs rely on something that the semiconductor manufacturers do not have to make. But the UTB wafers cost extra money—several hundred dollars more—so it’s going to add to the cost. There may be savings in device fabrication cost, which is difficult to estimate.
I think some companies, clearly Intel and some larger companies, will go with FinFETs, and some smaller companies will go with UTBs. Once both are in production, then people will be able to compare the benefits side by side very easily—the economics as well as the performance.
IEEE Spectrum: Does one offer a clear advantage over the other?
Chenming Hu: Staying with the hose analogy, when you have a big thick hose, you can carry more current, so it’s good for high speed. So that’s why I think the large companies that can make the investment in FinFETs will probably do it, because FinFETs are versatile. For the companies that need to have a quick way to get beyond 22 nm, I think UTB is a viable technology, especially for those companies that already have experience with the SOI. The shortcoming, of course, is the flat hose—with such thin silicon, less current goes through, which translates to lower speed.
IEEE Spectrum: Some industry leaders have been quoted as saying FinFETs won’t be as good for low-power applications as UTBs.
Chenming Hu: I think FinFETs are good for low power as well, but it does take more investment to bring them to production. I think a healthy competition will ensue. The fact is, the two technologies could coexist for a while, which is a good thing for the industry and certainly a very good thing for consumers.
Rachel Courtland, an unabashed astronomy aficionado, is a former senior associate editor at Spectrum. She now works in the editorial department at Nature. At Spectrum, she wrote about a variety of engineering efforts, including the quest for energy-producing fusion at the National Ignition Facility and the hunt for dark matter using an ultraquiet radio receiver. In 2014, she received a Neal Award for her feature on shrinking transistors and how the semiconductor industry talks about the challenge.